ECLIPSE-Net: an event-conditioned log-state estimator for remaining useful life prediction of lithium-ion batteries
Measurement Science and Technology · 13.11.2026 · 37 (45) · 456106
DOI: 10.1088/1361-6501/aeb8a6 ↗Kurz gefasst
Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 11.11.2026. Verfasst von Zhiyuan Xu und 1 weiteren Autor:innen. Die Arbeit nennt 51 Literaturangaben.
Abstract (Originalsprache)
Remaining useful life (RUL) prediction of lithium-ion batteries is a key measurement task for electric vehicles, aerospace and grid storage. Capacity fading is nonlinear and non-stationary; treated as a generic time series it loses its degradation-specific signals, and a single linear head cannot balance trend extrapolation, state-driven prediction and conservative estimation near the end-of-life threshold. This paper proposes ECLIPSE-Net, an event-conditioned log-state estimator combining a degradation event tokenizer, a log-time memory lattice, a selective second-order MIMO state module, an energy evidence mixer and a robust head. Five repeated experiments on CALCE, NASA and TJU give an overall root-mean-square error of 0.009552 Ah, an aggregate 22.85% below the strongest baseline, and, when the capacity window is advanced by each new measurement, a mean absolute RUL error of 0.78 cycles. A Friedman test rejects equal accuracy and ECLIPSE-Net ranks first in every test cell, but with the cell as the unit of analysis no pairwise comparison is resolvable, so the reduction is reported as an aggregate magnitude and not as a per-baseline significant difference. A complete two-level five-factor ablation attributes the result to the anchored bounded increment and its interaction with the second-order state module; the other three modules are not separable from run-to-run noise there. Trained over the whole benchmark, a model without them is worse in all six cells, by 1.8% of RMSE, at 60% fewer parameters. Deep-ensemble intervals cover 97% of observations at a nominal 95% level under leave-one-battery-out calibration. A thirty-configuration transfer study finds the model ahead in only ten, identifying the reason: its per-cycle event channels require degradation resolvable above measurement noise. Three rival explanations were rejected; subsampling the training cycles until that precondition holds improves the failing transfers six to twelve fold, leaving a recurrent baseline unimproved.
Zitieren
APA
Xu, Z., & Chen, J. (2026). ECLIPSE-Net: an event-conditioned log-state estimator for remaining useful life prediction of lithium-ion batteries. Measurement Science and Technology, 37(45), 456106. https://doi.org/10.1088/1361-6501/aeb8a6
Vancouver
Xu Z, Chen J. ECLIPSE-Net: an event-conditioned log-state estimator for remaining useful life prediction of lithium-ion batteries. Measurement Science and Technology. 2026;37(45):456106. doi:10.1088/1361-6501/aeb8a6
BibTeX
@article{Xu2026,
author = {Xu, Zhiyuan and Chen, Jianfei},
title = {ECLIPSE-Net: an event-conditioned log-state estimator for remaining useful life prediction of lithium-ion batteries},
journal = {Measurement Science and Technology},
year = {2026},
volume = {37},
number = {45},
pages = {456106},
doi = {10.1088/1361-6501/aeb8a6}
}
EUVIMED